Speculative Forcing speeds up autoregressive video generation
An arXiv paper introduces Speculative Forcing, a framework for few-step autoregressive video generation that treats intermediate denoising estimates as drafts and predicts which tokens can be accepted early and frozen as final output, while their cached keys and values stay available as context for remaining tokens. The authors explore two verifier designs: a training-free one based on similarity to clean content and stability across steps, and a learned one that reads the frozen generator's intermediate features to estimate each token's remaining gap to its full-step output. They report matching the four-step generator on LongLive with about 1.7 denoising steps per token and over 1.5x faster running at 41.7 and 37.8 fps, and preserving quality on Lingbot-World 2.0 with about 2.5 steps per token and 1.35x faster at 24.5 fps.
Cutting per-token denoising work without losing quality could lower inference cost for streaming video and world-model generation.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2610.22347 (cs)
[Submitted on 15 Oct 2026]
Title:Speculative Forcing for Efficient Autoregressive Video Generation
Authors:Hyunwook Choi, Seungryong Kim, Dongyoon Han, Sangdoo Yun
View a PDF of the paper titled Speculative Forcing for Efficient Autoregressive Video Generation, by Hyunwook Choi and 3 other authors